GLiNER2.5 base, fine-tuned on Re-DocRED entity spans
Entity span finder for the edgextract product path. Trained on Re-DocRED train mention spans only (PER, ORG, LOC, TIME, NUM, MISC). No relation head.
Base weights: fastino/gliner2.5-base-v1. Encoder backbone in the config is DeBERTa-v3.
How to use
Load with GLiNER2 (gliner2). Inference threshold used in the kept product row: 0.40.
This checkpoint finds names. Relation scores in edgextract still come from ATLOP (raphaelmansuy/atlop-roberta-base-redocred), not from this model.
Measured (Re-DocRED dev, edgextract)
| Mention F1 | 0.813 (precision 0.832, recall 0.794) |
| Product relation Ign F1 with this finder + ATLOP + alias alignment | 63.57 |
| Same, plus Mistral Small on the uncertain band | 64.37 |
Official gold-entity ATLOP Ign F1 on the same dev set is 76.81. The gap is the product path (found names), not this file alone.
Training
Re-DocRED train documents only. Dev was used to pick the checkpoint and threshold, not as a training label. Test was not opened for this fine-tune.